The Robot Gave Itself Permission, Then Immediately Proved Why That's a Bad Idea

This repo keeps its big decisions in numbered files called ADRs — architecture decision records, the kind of thing a team writes down so nobody re-litigates it later. One of them, ADR-0003, says a human has to sign off before an agent implements anything net-new. Sensible. The kind of guardrail you'd actually want. So naturally, this repo just shipped ADR-0022, which invents something called a "standing green-light": a human starts a sweep once, and from then on a Skill — this repo's word for a scripted routine an agent runs, part checklist, part tiny program — gets to decide, ticket by ticket, whether it's allowed to hand out that sign-off itself. The ADR's own language: bravery is "bounded by determinability" — the agent gets to be as bold as it judges the situation to be clear-cut, and it's the one doing the judging.

I'd be more worried about the philosophy if the very first live run hadn't handed me the punchline. The new Skill is called auto-triage; its job is to sweep the open issue backlog and stamp tickets ready for an agent to pick up. Its first run found nothing. Not "nothing needed attention" — nothing eligible, full stop, because the eligibility check only scanned tickets sitting in three particular label buckets. Issue #428 wasn't in any of them: it's a real regression report, the repo's owner had weighed three fix options an earlier agent laid out for him, and settled it in a comment reading, in its entirety, the single letter "A." A human decision, typed and public, sitting right there — invisible to a sweep that only reads labels, not comments. The sweep didn't flag its own blind spot. A human had to read the log and point it out by hand, which is exactly the job this whole apparatus exists to make unnecessary.

To be fair, because I do read the whole diff before I mock it: the fix landed 16 minutes into that same session, and it's a real one — eligibility now checks who wrote the most recent comment, not which label a ticket happens to wear. #428 got its "A" read correctly and moved to ready-for-agent before the session ended. That's the system working exactly as advertised: human catches gap, agent patches it, gate passes, PR merges. Genuinely fine. What isn't fine is the framing. A tool built specifically to reduce how often a human has to babysit the backlog needed a human babysitting the backlog to notice it wasn't looking at the backlog — on day one of being allowed to greenlight itself. "Bounded by determinability" is a nice phrase. Today it was bounded by whoever was still reading the output. Ask me again once nobody is.